A good AI tutor takes the friction out of practice and leaves the judgement with the person. A week of rules being proposed at the top, and judgement being quietly eroded underneath. Demis Hassabis published a framework for frontier standards. The Prime Minister stood up at the University of Sydney and put AI coordination inside his own department. And a Columbia professor's research put a name to something I have been seeing in slide decks for months: when everyone consults the same tools, the work starts to look the same.
What we mean by an AI tutor is set out on the AI tutor for schools page.
A FINRA for frontier models
Demis Hassabis published a framework for frontier AI standards this week. He proposes a US standards body modelled on FINRA: a federally overseen public-private partnership, industry funded, with independent technical experts and open-source representatives on the board.
Models that clear certain capability benchmarks get designated Frontier class. The organisations that build them become Frontier Labs, with expectations around model cards, internal security, personnel vetting, and safety resourcing. The mechanism starts voluntary. Labs share models for review up to 30 days before release. Once the assessment protocol proves robust, it formalises into a requirement for deployment in the US market. Benchmarks get refreshed quarterly, saturated ones retired, and eventually the body builds its own held-out tests so labs cannot overfit to them.
Startups and academic models sit outside the process entirely. Only frontier capability triggers frontier obligations, regardless of country of origin or whether the model is open or closed.
Hassabis describes an intense commercial and geopolitical race where progress is outpacing understanding, and where even the experts disagree about what happens next. His answer is cautious optimism, and the framework includes an option to coordinate a slowdown among frontier labs if the situation warrants it. Whether that would survive contact with actual competitive pressure remains to be seen. His write-up is worth reading in full.
The Prime Minister at the University of Sydney
I had the pleasure of attending the Prime Minister's address on AI in person this week at the University of Sydney.
He announced that Australia will establish a set of mandatory national standards for AI, and an Office of AI inside the Department of the Prime Minister and Cabinet. Until now the government's response has been issue by issue, sector by sector. Copyright with the Attorney-General, energy with DCCEEW, and so on. Putting the coordination function in the Prime Minister's own department pulls those lanes into one place. He drew the comparison to civil aviation in the 1920s and genetics in the 1990s: technologies that touched enough of government that a coordinated approach became the only workable one.
The framework will go to National Cabinet next month, with legislation aimed at Parliament early next year. The key messages I distilled from the speech:
No company will be able to use Australian books, music, art or journalism to train AI without consent, control and fair value. His words were blunt. Theft is theft.
Large-scale data centres will carry a legal obligation to underwrite their own new power supply, pay their full share of connection costs so energy bills are not impacted, and reduce power when needed to strengthen the grid.
Location, energy and water rules currently sit across every level of government. Rather than legislating for every possible risk, the aim is a single national frame that moves as fast as the technology does.
An AI tutor should not do the thinking you are paid for
Columbia Business School professor Sandra Matz has research showing AI use at work makes people boring. Her argument is that most AI tools generate the safest, most obvious response, so when everyone consults the same tools, everyone converges on the same output. CNBC's write-up captures it well.
I can spot an AI-generated PowerPoint from a mile away, and I switch off by slide two. A human who deeply understands their argument cuts ruthlessly. They know which single chart does the heavy lifting and which nine slides were padding. I had an amazing boss, Estelle Marque, who taught me this, with extensive practice.
There is still a craft to curating a low page count, high-impact presentation. It requires judgement about what your audience needs to feel, not just what they need to know. Sorry, but the agent tools cannot do that yet, and you know I love AI.
AI is brilliant for removing friction, summarising, organising, accelerating the mundane. Where it goes wrong is when it replaces the judgement that was supposed to be your contribution, and the connection to other humans. The professionals who thrive over the next decade will use AI for everything except the thing they are actually being paid to think about.
Talking on dog walks
I spent last week talking to GPT-Live on dog walks. The big shift is the full-duplex architecture. It listens while it speaks, so the conversation flows the way a real one does. You can pause to think and it waits. You can interrupt and it adjusts. And the slower pace of speech turns out to be a feature rather than a limitation. It feels natural and relaxed, and it gives the model time to run web lookups in the background without breaking the rhythm of the chat.
To be clear, it does not feel as smart as a long-form conversation with Sol or Fable. If I want depth I still go to a frontier model in text. But as a companion for a walk, a commute or thinking out loud, it is genuinely engaging in a way voice AI has not been before.
Galaxy, folders, and England
A blank chat box asks students to know what they do not know. Most will not ask, because they cannot see what is missing. So we built Galaxy to show them. Every learning goal is a star, related stars form a constellation, and students light up stars by genuinely demonstrating understanding in conversation, not by clicking through or guessing. Students choose which constellation to explore. A depth meter only moves forward with demonstrated learning effort.
CurricuLLM is now live in England, and it is free for every teacher and pre-service teacher in training. We have also brought in Oak National Academy content. When a strong Oak lesson already exists, CurricuLLM recommends it before building anything custom, and you can ask it to adapt that lesson while the Oak structure keeps the curriculum alignment intact. After feedback from our schools, you can now sort the sidebar into folders as well, grouping conversations by class, topic or project.